Tom Head’s Post

🤔 Struggling to get all the results you want from your RAG app? Here’s how to fix it! When you have a large collection of documents, getting all the results you need can be challenging. Here’s a common scenario: 1️⃣ You have an index with thousands of documents on similar topics.  2️⃣ You query a specific topic that is mentioned in several documents, but only a few focus extensively on it.  3️⃣ Chunks from the few focused documents dominate your results, even though you want to see relevant information from all the documents. 𝗪𝗲 𝗿𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱 𝘀𝗲𝘁𝘁𝗶𝗻𝗴 𝗮 𝗵𝗶𝗴𝗵 𝗞 𝘃𝗮𝗹𝘂𝗲 𝘁𝗼 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗲 𝗯𝗿𝗼𝗮𝗱𝗲𝗿 𝗿𝗲𝘀𝘂𝗹𝘁𝘀. However, this approach can introduce noise. To refine your results, you can use a reranker to cut off the low-ranking results and increase precision. This is how you would do it from scratch, which requires experimentation to find what works best. Keep in mind that this will increase latency. With trufflepig, you don't need to compromise on latency to get the results you need. 👋 Have you faced similar challenges? Share your thoughts in the comments and follow trufflepig for more AI content 🐷 Try trufflepig: (https://lnkd.in/gdW98A38)

We've been teasing our first Corso AI feature - it's now here for everybody 🎉 Inside the Corso app, you can now query your data by asking it questions! In the video we ask "What was the SKU of my most returned product?", and we get the answer back (with the number returns also given, as a bonus). Maybe you want to constrain it to this month. The AI knows about that too, and can answer a question like "What was the SKU of my most returned product this month?" If you already use Corso, try things like "By month, how many orders had a shipping protection claim in the last year?" to see a breakdown. Or: "What was the most commonly return reason for my 'Gray Stonewashed Jeans' Product in the last 6 weeks?" AI stands to make our software more powerful, and easier to use. Try it out, and let us know what you think. The team had a ton of fun building this, we hope you will have just as much fun using it!

When analyzing performance of some part of your company, do you prefer: a) a written explanation only b) charts and numbers only c) both Personally, I'm a "c: both" guy. But if I can't have both, I'm fine with just "b: charts and numbers." We're about to roll out a new feature in the Databox Growth plan that uses generative AI to summarize the performance of all of the metrics on a dashboard. See image below for an example. What's novel about this is that no one needs to write this. It's written for our customers. That obviously wasn't that feasible before generative AI. Customers have reacted very positively to our generative AI summaries in other parts of the app. So, we're expanding our rollout. Dashboards are our most popular feature. So, this is the obvious place to put it to maximize value for our customers. If your answer is "a: a written explanation" or "c: both" what do you think of this implementation?

It's not enough to just be "hyped" about GenAI. How much is this thing actually going to COST for your organization? Sharing below one of the best clear and practical posts I have seen about Generative AI in recent weeks. Check out both posts referenced here on Costs associated with RAG LLM apps and how to reduce them. Definitely worth a look if you are exploring building RAG apps at your organization.

At Fix AI, our AI Agents can hold very natural, human-like conversations with people. We use functions to lead the conversations down different paths, allowing our AI Agent to make decisions. Here's some common misconceptions and an example function.

What does the release of`gpt-4o-mini` mean for those developing AI apps? 👉 This from the perspective of my text-mining app which helps you find insights from user reviews, customer surveys, etc. =============== What's the big deal? =============== ✅ gpt-4o-mini is a FASTER version of gpt-4o. ✅ gpt-4o-mini is CHEAPER than gpt-3.5-turbo (60% cheaper!). ❌ gpt-4o-mini is not as good as gpt-4o for complex tasks. ================== So, what does this mean? ================== 💡 Where my app was using 3.5-turbo, switch to 4o-mini 💡 Where my app uses 4o--Can make the switch to benefit from performance gains without impact to accuracy. ============================ THE biggest benefit of the new model ============================ Reflective prompting just got a WHOLE lot cheaper! ℹ Reflective prompting is where you get the LLM to double-check its own output. This improves accuracy! My view: The price drop and performance improvements means that reflective prompting has now a major part of AI development. But a lot of testing needs to be done to validate accuracy using reflective prompts. 🙄 My plans for today has suddenly changed. But, so far my testing with reflective prompting is looking very promising indeed! (I know, I'm not talking much about experimentation lately, but I will get back to it, I promise! 😂 ) #ai #userresearch #cro

One core skill in the Gen AI era is often overlooked. And it's not difficult or too technical! You just have to... Start taking advantage of cheaper and faster LLMs! ⠀ Thanks to GPT-4o Mini and Claude 3.5 Sonnet... And some powerful no-code AI app builder solutions... We can now do incredible things with prompt chaining! ⠀ In my Masterclass on prompt chaining & meta-prompting... 𝗜'𝗺 𝗴𝗼𝗶𝗻𝗴 𝘁𝗼 𝘀𝗵𝗮𝗿𝗲 𝘁𝗵𝗲 𝗳𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴: ✅ 7 high-value prompt chain workflows ✅ How to build prompt-chain tools on 4 no-code platforms ✅ How to inject rich-context dynamically into your prompt chains ✅ Give you the full prompts and prompt resources ✅ Master meta-prompting techniques to enhance prompts ✅ How to perform conditional logic on prompt chains ✅ How to use Route LLM techniques to save on API costs ⠀ 🎁 𝗚𝗲𝘁 𝟭 𝗼𝗳 𝘁𝗵𝗲 𝗵𝗶𝗴𝗵-𝘃𝗮𝗹𝘂𝗲 𝗽𝗿𝗼𝗺𝗽𝘁 𝗰𝗵𝗮𝗶𝗻𝘀 𝗳𝗼𝗿 𝗙𝗥𝗘𝗘 𝗿𝗶𝗴𝗵𝘁 𝗻𝗼𝘄: Comment "𝗣𝗿𝗼𝗺𝗽𝘁 𝗖𝗵𝗮𝗶𝗻" now — I'll send the 6 prompt workflow. ⠀ Add these incredible AI skills to your toolbelt! Join me on Tomorrow, 8/10, at 12 PM ET / 9 AM PDT 🔴 LIVE! Click the "𝗩𝗶𝘀𝗶𝘁 𝗺𝘆 𝘄𝗲𝗯𝘀𝗶𝘁𝗲" link above to register now.

An AI cursor that clicks, navigates, and gets stuff done for you??? The Browser Company is teasing a browser with that kind of AI magic. At Orango AI, we can do it TODAY—and you don’t even need a new browser. Our AI isn’t general-purpose (we didn’t raise $68M, sorry), but if your product has an Orango AI Assistant embedded, it can do any task within your app for your users. Have a complex product? Want to guide your users with ease? Let’s chat! Btw Josh Miller, your ADK looks awesome (and I love Arc). 😍 Let’s connect! P.S. video of the screen > screen recording 😅

Here are my top best practices for building production-ready AI apps: 1. Build evals (define test cases to ensure you're actively improving your app & not causing any regressions) 2. Break down one LLM call into multiple (AI systems do a lot better when you have many LLM calls chained together. i.e, instead of sending an LLM call to a model to generate code, send it to a "architect" model to generate a plan, then a "coding" model to generate code, then a "reviewer" model to verify) 3. Start simple (with 1 LLM call), then iterate with prompt engineering (few shot examples, chain of thought, descriptive prompts) before building a more complex system with chained LLM calls 4. When dealing with data you need the LLM to use as context, use RAG (then you can iterate with your evals to try different chunking/embedding strategies, adding a re-ranker, ect..) 5. When you want to have the LLM optimize on a domain-specific thing or on a specific style (i.e, write emails like you do), use finetuning 6. When launching, ship with observability and look at the data. It's so important to look at how people are using your system, then you can do things like segment customer prompts, run evals on them, and get great info on where you need to improve your AI system Going to make a video on this next week, including my favorite tools to use for each of these steps!

Using AI assistant functions can be a game-changer—if you know how! Here's how to truly harness their power and avoid common mistakes. Misconceptions: 1. Some people think that if you tell the model what your function does, it will just perform that function. Wrong. LLMs can only output tokens, and functions are no exception. Assistants are really just generating the ARGUMENTS to a function. 2. Using functions can cut down on token costs. Wrong again. Your assistant needs to know which functions are available, so every function schema adds to the input tokens. 3. Functions vs response format. While these two features are similar, they serve distinct purposes. Specify a response format when you need your assistant's outputs to have a consistent structure. Use functions when you want your assistant to generate inputs and interpret outputs from your app. Function schemas also come with parameters that can be confusing. Here’s a quick breakdown: "additionalProperties" - Set this to true if you’re okay with the model including extra parameters you didn’t specify. "strict" - Ensures the model sticks exactly to your schema. Unless you're experimenting, this should be set to true. "parallel_tool_calls" - Enable this to allow multiple functions in a single run. "tool_choice" - Options include: auto (letting the assistant decide when to call a function), required (the function must be called every time), or none (no functions are used). Feeling overwhelmed? Check out one of the functions we've been using in the image below! This allows your assistant to pause the conversation for a specified reason.